Autonomous Driving

Today's Stories

1. 1

   DrivingBench Sandbox rename got AI to drive a real Corolla

   DrivingBench, run by three engineers at the AI startup Axiom Math, connected GPT-6 Astra, Claude Fable 5.1, Grok 4.6 and GPT-5.6 Sol to a 2022 Toyota Corolla. Only GPT-6 Astra completed the roughly 135 m course, in 5 minutes 22 seconds. Drivers were told to brake at any time and the vehicle was speed-limited through a cone-lined course, so the result is a benchmark of physical capability, not a usable self-driving system.

   The team is considering DrivingBench v2, with more trials per model and harder courses, so whether the outcome holds depends on whether the renamed sandbox still works at greater scale.

2. 2

   Turing autonomous drive feels natural, but AI unreadable

   A test ride in Turing's self-driving car showed natural-feeling driving, yet there were moments the AI's thinking could not be read. The ride suggests Turing's autonomous driving may be smooth in practice, but its AI's decision-making appears hard to follow, which could complicate trust.

   Whether Turing can make its AI's reasoning understandable will likely shape how riders and regulators judge it.

3. 3

   Desynced sensor feeds can silently corrupt VLA and ADAS models

   The article uses an Xpeng owner's Reddit complaints — Spotify streaming dropouts, mistimed ADAS alerts, and a cabin thermometer reading 2 degrees off — as analogies for cross-modal misalignment in VLA training data. Each modality can pass its own quality check while sync errors slip through, so the model absorbs the contradiction into its weights and stops reflecting real-world physics, the article says.

   The piece suggests source-level synchronization, not post-hoc alignment, is the fix, but the checklist is self-assessment guidance rather than a proven fix, and it ends with a Bright Data free trial offer for engineers.

4. 4

   Microsoft merges Office and coding into a single Copilot

   Microsoft is consolidating Home, Code, and an OpenClaw-based Autopilot into Copilot. Autopilot's agents are an enterprise product that executes tasks unattended, raising the possibility that routine clerical and development work shifts to AI agents.

   Home and Code are only slated for testing in Frontier, while Autopilot is unscheduled until the end of September 2026 and unproven.

5. 5

   Tesla's Ara bot forecast: FSD V15, Unsupervised driving soon

   CleanTechnica writer Fritz Hasler says Tesla's in-car Grok bot, Ara, offered an unprompted forecast that FSD V15 will be released at the end of this year or the beginning of next, with Unsupervised Full Self Driving. If that forecast holds, a driver-assistance suite the author already calls essentially flawless on highways could arrive without a person at the wheel, which he sees as a major step for long trips.

   The forecast is the bot's, not a Tesla announcement, and the author doubts V15 lands before next year — so the test is whether Unsupervised FSD is limited to highways or elsewhere. Watch the $99/month price.

6. 6

   Archer's Boeing unit deal clears Hart-Scott-Rodino wait

   Archer Aviation's planned purchase of Boeing's Wisk Aero, SkyGrid and Insitu units passed the Hart-Scott-Rodino waiting period, which expired at 11:59 p.m. EDT on Sept. 18. The deal would fold autonomous flight systems, airspace software and uncrewed aircraft into Archer's platform for aerospace and defense.

   The deal still needs other regulatory approvals and customary closing conditions; Archer expects it to close by the end of 2026.